Deep learning found 31.9% of one Indonesian regency's standing water held mosquito larvae

A health team in Pangandaran flew a consumer drone over a single Indonesian regency, fed the photos to a deep-learning model, and let it pick out every puddle in 4,400 images. Every one of the 47 standing-water sites the model flagged actually held water in the field. Fifteen of them, 31.9%, held mosquito larvae, including the two primary local malaria vectors. The result, published this month in Scientific Reports, is the cleanest low-cost playbook for mosquito habitat mapping that 2026 has produced, and it runs without the satellite, multispectral sensor or image-stitching pipeline that have kept the technology out of reach of district health offices.

What the paper actually did

The Pangandaran Public Health Laboratory, a unit of Indonesia's Ministry of Health, partnered with researchers from Ehime University (Japan), Lurio University (Mozambique), Tokyo Women's Medical University and Universitas Padjadjaran. They asked: can a deep-learning model find standing water in cheap, unprocessed drone images, and skip the expensive step of stitching them into a single georeferenced map?

The answer is yes. The team flew consumer drones over Pangandaran Regency, a coastal district on Java's southern edge, and collected more than 4,400 still images in colour and grayscale, each with its own GPS tag. The team ran them through a DeepLabV3+ model with an EfficientNetV2 backbone, a workhorse of low-cost computer vision in 2026.

The model did not try to stitch the photos into a mosaic. It scored each image individually, used the GPS metadata to place positive predictions on a map, and pushed everything through a cloud-based pipeline. No specialised hardware. No proprietary image-stitching software. No multispectral sensor.

What the model found

The segmentation quality, measured by mean Intersection over Union (mIoU), was 0.86 on the colour images and 0.80 on grayscale, both publication-grade scores for water-body segmentation in aerial imagery.

The number that matters more for a public-health audience is what happened in the field. The team visited 47 sites the model had flagged as standing water. Every one held water. Fifteen of the 47, 31.9%, contained mosquito larvae, including the two principal local malaria vectors Anopheles vagus and Anopheles sundaicus, the latter a brackish-water breeder that thrives in the coastal lagoons of southern Java.

The 31.9% figure is the editorial point. Most standing water does not contain mosquito larvae; the model does what a human surveyor would do, flagging every puddle, ditch and rice-paddy corner, and letting a second pass sort the real larval habitats from the irrelevant ones. The paper's innovation is the first pass: making it cheap, fast and scalable.

Why the "no orthomosaic" bit matters

Until now, drone-based habitat mapping has meant producing an orthomosaic, a single, high-resolution, georeferenced image stitched from hundreds of overlapping photos. The step requires commercial software (Pix4D, Agisoft, DroneDeploy, annual licences in the low four figures), a desktop workstation, and hours of processing per survey. For a research group at a well-funded university, this is routine. For a district health office in a malaria-endemic region, it has been a non-starter.

Francisco and his colleagues skipped it. They used the GPS coordinates embedded in each image, and let the cloud do the heavy lifting.

The three things that are converging in 2026

The Pangandaran paper lands in the same week as three other peer-reviewed results that change what "vector control" means for a tight-budget health office.

The atmospheric-and-urban-form literature has just closed a long-standing loop. The 18 June PNAS paper by Lugão and colleagues at the Federal University of Juiz de Fora and the Universidade Federal de Goiás models Aedes aegypti populations across Brazilian cities with both atmospheric and urban-form covariates, and finds that urban morphology is a stronger driver of hotspots than temperature alone. The 19 June iScience paper by Liu and colleagues extends the logic to 108 countries, with flood duration as the driver. The mosquito is a city-scale problem before it is a weather-scale problem.

The community-prevention literature has produced its cleanest datapoint. The Philippines Department of Health reported 50,727 dengue cases in the first five months of 2026, a 56% decrease on 2025, and credited the "4Ts" campaign: Taob, Taktak, Tuyo, Takip (empty and overturn containers, shake out water, keep surroundings dry, cover water containers). The 4Ts work, but only when a community knows where the breeding sites are. The Pangandaran paper is the missing upstream piece.

The AI-and-imagery literature has now produced its first field-validated low-cost workflow for larval habitat mapping. The Pangandaran paper is the first peer-reviewed demonstration that the workflow holds together in a malaria-endemic region, with a local vector-control team in the field, and without the infrastructure that has historically kept the technology in the rich-world research literature.

The three prongs do not replace each other. The atmospheric models tell a city where its hotspots will be; the community campaigns tell a neighbourhood what to do; the drone-and-AI workflow tells a field team which puddles to empty first.

The 31.9% finding is a useful reminder at the household level: most standing water is not a breeding site. The personal protection layer does not change: empty saucers, turn over buckets, cover water-storage containers, sleep under treated netting or in screened rooms, and use a proven repellent on exposed skin at dusk and dawn.

What to do

For a household or small site, the practical takeaway from the Pangandaran finding is unchanged: most standing water is not a breeding site, so the work is in finding the few that are.

• Empty or overturn any container that can hold water for more than a few days (buckets, plant saucers, tarpaulins, old tyres).
• Shake out, scrub and re-fill pet drinking bowls and bird baths at least weekly; larvae take roughly seven to ten days to mature.
• Cover water-storage tanks and barrels with tight-fitting lids or fine mesh.
• Clear roof gutters and flat-roof drains before the rainy season.
• Fill or drain low spots in the garden, and keep fish in ornamental ponds where feasible; larvivorous fish suppress Culex and Anopheles breeding.
• Use proven personal protection at dusk and dawn: long sleeves and trousers, a treated net or screened room, and a repellent on exposed skin.
• For district or municipal teams, the Pangandaran workflow is now the strongest published case for adding low-cost drone surveys to routine larval-source management, especially during the rainy season.

What to watch in the next twelve months

Validation in other regions. The Pangandaran paper is a single district in coastal Indonesia. The workflow will need re-running in a Sahelian country, a South American city (where Aedes aegypti dominates) and a Pacific island (where the human-wildlife interface is the active frontier). The architecture is portable; the training data is not.

An open-source pipeline. The Francisco team used DeepLabV3+ and EfficientNetV2, but the model weights and pre-processing code are not yet public. The most consequential follow-up would be a public release of a pre-trained model any health department can download.

The integration question. The most useful next paper chains the Pangandaran workflow to a vector-control decision: model finds a site, field team gets a phone notification, worker confirms the larvae, the local 4Ts-equivalent campaign deploys within 48 hours. That end-to-end loop is the operationally interesting thing. The Pangandaran paper is the first piece.

What we know

• A deep-learning model trained on 4,400 still drone images from Pangandaran Regency, Indonesia, identified standing-water sites with a mean Intersection over Union (mIoU) of 0.86 on colour images and 0.80 on grayscale, a publication-grade score for water-body segmentation in aerial imagery. Francisco et al., Sci Rep (2026)
• Field validation of 47 sites the model flagged as standing water confirmed water presence in 100% of cases; 15 of those sites (31.9%) contained mosquito larvae, including the primary local malaria vectors Anopheles vagus and An. sundaicus. Francisco et al., Sci Rep (2026)
• The workflow bypasses orthomosaic generation entirely. It uses the GPS metadata embedded in each individual drone image to place the model's positive predictions on a map, runs in the cloud, and requires no specialised hardware or proprietary image-stitching software. Francisco et al., Sci Rep (2026)
• The architecture is DeepLabV3+ with an EfficientNetV2 backbone. The study was led by researchers from Ehime University, Lurio University, the Pangandaran Public Health Laboratory (Indonesia's Ministry of Health), Tokyo Women's Medical University and Universitas Padjadjaran. PubMed 42315628
• The work was funded by the Japan Society for the Promotion of Science (JSPS) Joint Research Project grant JPJSCCB20240008, and the paper is published open-access in Scientific Reports (DOI 10.1038/s41598-026-58240-4). Francisco et al., Sci Rep (2026)

Sources cited

1. Francisco, Micanaldo Ernesto, Andri Ruliansyah, Firda Yanuar Pradani, Gaku Masuda, Lia Faridah, Kozo Watanabe. "Deep learning identifies water bodies from low-cost drone images for mosquito larval habitat mapping." Scientific Reports (Nature), 18 June 2026. DOI 10.1038/s41598-026-58240-4. Open access. https://www.nature.com/articles/s41598-026-58240-4
2. Francisco et al. (2026), PubMed record, PMID 42315628, indexed 18 June 2026. https://pubmed.ncbi.nlm.nih.gov/42315628/
3. Liu, Q., Zhang, S., Liu, M., Liu, J. "Impact of flood duration on dengue burden across 108 countries." iScience (Cell Press), eCollection 19 June 2026 (online 25 April 2026). DOI 10.1016/j.isci.2026.115917. https://pubmed.ncbi.nlm.nih.gov/42169817/
4. Lugão, P. H. G., da Silva, M. R., Cascelli, R., Chapiro, G. "Spatial and temporal prediction of Aedes aegypti populations with atmospheric and urban forms dependence." PNAS 123(25):e2533964123, published online 18 June 2026. DOI 10.1073/pnas.2533964123. https://pubmed.ncbi.nlm.nih.gov/42313935/